README for SHRiMP2: SHort Read Mapping Package, version 2.0
http://compbio.cs.toronto.edu/shrimp

This document describes  the programs of which  SHRiMP2  is comprised, including
their use, output  formats, and various parameters.  The algorithms employed are
also    briefly  described. SHRiMP2 builds  on,    and significantly extends the
original SHRiMP package, henceforth referred to as SHRiMP1.

SHRiMP2 is brought to you by:
	Matei David
	Michael Brudno
	Daniel Lister
	Misko (Michael) Dzamba

SHRiMP1 was originally developed by:
	Stephen M. Rumble
	Michael Brudno
	Phil Lacroute
	Vladimir Yanovsky
	Marc Fiume
	Adrian Dalca
	Matei David

For details on the original SHRiMP1 package, see the following publication:

    Rumble SM, Lacroute P, Dalca AV, Fiume M, Sidow A, et al.
    (2009) SHRiMP: Accurate Mapping of Short Color-space Reads.
    PLoS Comput Biol 5(5): e1000386. doi:10.1371/journal.pcbi.1000386

The authors may be contacted at shrimp at cs dot toronto dot edu.


--------------------------------------------------------------------------------
Table of Contents
--------------------------------------------------------------------------------

  1. Overview

  2. Minimum Requirements

     2.1 RAM Requirement
     2.2 Other Requirements

  3. Sample Usage Scenarios

     3.1 Installing SHRiMP2
     3.2 Compiling SHRiMP2 from source
     3.3 Mapping against a genome whose projection DOES fit in RAM
     3.4 Mapping against a genome whose projection DOES NOT fit in RAM
     3.5 Mapping cDNA reads against a miRNA database
     3.6 Advanced: Mapping in OVERLY sensitive mode

  4. Input and Output Specifications

  5. Parameters

  6. Algorithms

  7. Comparison to SHRiMP1

     7.1 Algorithms
     7.2 SHRiMP Output Format
     7.3 SHRiMP1 Tools

  8. Contact

  9. Acknowledgements


--------------------------------------------------------------------------------
1. Overview
--------------------------------------------------------------------------------

SHRiMP2 is a  software package for  mapping reads from a  donor genome against a
target (reference) genome. SHRiMP2  was primarily developed  to work  with short
reads produced by  Next Generation Sequencing (NGS)  machines. The main features
of SHRiMP2 include:

  - Fasta input format;

  - SAM output format;

  - Support for both  letter space (Illumina/Solexa) and colour space (AB SOLiD)
    reads;

  - Paired mapping mode, using mate-pair/pair-end information;

  - Parallel  computation,  fully utilizing  the   power of modern shared-memory
    multi-core architectures.

SHRiMP2 uses two techniques to match reads to a genome: we  first use the q-gram
filtering  technique, utilizing  multiple spaced    seeds, to rapidly   identify
candidate  mapping locations for each  read.  Subsequently,  these locations are
thoroughly  investigated by a   vectored implementation of  the standard dynamic
programming  (Smith-Waterman)    string matching algorithm,   allowing  for  the
accurate   identification  of  mismatches  (SNPs), micro-indels,  and crossovers
(colour space sequencing errors.)

SHRiMP2 primarily targets sensitivity, but it has come a long way to achieve it
at a reasonable speed. To illustrate its performance:

  Sensitivity: SHRiMP2 achieves 94.4% precision  and  78.6% recall when  mapping
  simulated colour space reads of 50 base pairs (bp) each containing 1 SNP _and_
  1 indel of size up to 5bp _and_ 4% per-base error rate.

  Speed: SHRiMP2 on a single  3.0GHz core with  16GB of RAM  can map 36bp colour
  space reads against the reference  human genome (hg18) at  the rate of 160,000
  reads per hour, and twice as fast  if using mate-pair/pair-end data. This rate
  is fully core-parallelizable, in the sense that a cluster  of 4 machines, each
  with 2 quad-core CPUs and 16GB of RAM, will map 4x2x4=32 times faster. This is
  about 10 million  paired reads per hour.  So,  250 million  36bp paired reads,
  equivalent  to a 3x coverage of  hg18, would  take  about 30  hours.  Prior to
  mapping, indexing the projection of hg18 into 4 pieces  that would each fit in
  16GB   RAM can be  done by   in 2 hours  by   running 4 different  machines in
  parallel.  (These are all rough estimates.)


SHRiMP2  consists  of  the  read  mapper  program  gmapper,  along with a series
of tools  and  scripts.   In this README file,  we primarily  describe  gmapper.
Descriptions of tools remaining from SHRiMP1 are available in  README.SHRiMP132.
Descriptions of various helper scripts are in the script directory utils/.


--------------------------------------------------------------------------------
2. Minimum Requirements
--------------------------------------------------------------------------------

2.1 RAM Requirement
-------------------

SHRiMP2 is designed to fully utilize physical  random-access memory (RAM). While
it is certainly possible to run it on swap (disk-mapped) memory, the performance
cost of doing that is rather prohibitive.

  RAM  Requirement: To map against a  set of contigs of total  length L (in bp),
  using K spaced  seeds of weight W,  SHRiMP2 needs the following minimum amount
  of RAM, in bytes:

    ( L x K x 4 ) + ( K x 4^{W or 12(**)} x (4 + sizeof(void *)) ) + 50,000,000

The third term represents working memory.

The second term varies with the weight W of the seeds  used.  With the default 4
seeds of weight 12,  on 64 bit-machines (with  8-byte pointers), this amounts to
0.75GB. (**) By  hashing kmers (see  -H parameter), the  exponent can be brought
down to 12, at the expense of some speed drop.

The first term generally dominates. E.g., with the default settings (K=4 seeds),
to map against the full hg18 (L=3*10^9 bp), the first term becomes  3*10^9 x 4 x
4 = 48GB of RAM.

To make it accessible to machines with smaller amounts  of RAM, SHRiMP2 provides
a mechanism to split a genome into several chunks,  each of which comes with the
overhead of terms 2 and 3. E.g., in our tests, we split hg18 into 4 chunks, each
using  about 13GB of RAM, which  can fit comfortably on  a  machine with 16GB of
RAM.

Currently,  SHRiMP2 does not  split individual  contigs.  (However, this  can be
achieved by hand, possibly at the cost of losing some  mappings in the region of
the break  point.) As a result,  the minimum amount  of RAM  required equals the
amount needed to index the  largest contig. E.g., in  the case of hg18, chr1 has
about 260,000bp, and the amount of RAM required to index it  with K spaced seeds
is K x 1.04 +  0.75 GB. (For a  cluster of nodes with 4GB  of RAM, we  recommend
using K=3 seeds rather than the default 4.)


2.2 Other Requirements
----------------------

SHRiMP2   uses     SSE2 ("vector")  machine instructions    to   achieve  a fast
implementation of  the   standard dynamic programming   (Smith-Waterman)  string
matching algorithm. E.g., it will not work on a non-x86/x86_64 architecture.

SHRiMP2 uses the OpenMP API to achieve multi-threaded operation.


--------------------------------------------------------------------------------
3. Sample Usage Scenarios
--------------------------------------------------------------------------------

3.1 Installing SHRiMP2
----------------------

Assume we downloaded the linux binary package in
./SHRiMP_2_1_1.lx26.x86_64.tar.gz

  $ tar xvzf SHRiMP_2_1_1.lx26.x86_64.tar.gz
  $ cd SHRiMP_2_1_1
  $ file bin/gmapper
  bin/gmapper: ELF 64-bit LSB executable, AMD x86-64, version 1 (SYSV), for
  GNU/Linux 2.6.9, statically linked, for GNU/Linux 2.6.9, not stripped
  $ export SHRIMP_FOLDER=$PWD

Done! At this point the binaries are in bin/, and the various scripts are in
utils/.


3.2 Compiling SHRiMP2 from source
---------------------------------

Assume we downloaded the source package in ./SHRiMP_2_1_1.src.tar.gz

  $ tar xvzf SHRiMP_2_1_1.src.tar.gz
  $ cd SHRiMP_2_1_1
  $ make
  [...warnings, hopefully no errors...]
  $ ls bin/gmapper*
  bin/gmapper bin/gmapper-cs bin/gmapper-ls
  $ export SHRIMP_FOLDER=$PWD


3.3 Mapping against a genome whose projection DOES fit in RAM
-------------------------------------------------------------

Assume we have:
./ciona.fa - the Ciona Savignyi genome (175*10^6 bp)
./reads.5m.50bp.cs.fa - 5 million 50bp colour space reads

We look at the genome and the reads:

  $ ls -lh ciona.fa
  -rw-r--r-- 1 username 173M May  6 14:23 ciona.fa
  $ head -n 4 ciona.fa
  >reftig_0
  ACCACACATCTGGCGTTGGTGCAACTACCACAACATCTGGCGTTCGTGCA
  ACCACCACAacatctggtcctggtgcaactaccacaacatctggtcctgg
  tgcaaccaccacaacatctggtgctggtgcaaccaccacaacatctggtg
  $ head -n 4 reads.5m.50bp.cs.fa
  >469_26_42_F3
  T121133100312321122210031200212212233
  >469_26_379_F3
  T312022230033100001303023233122232120

Assume we  have 2 machines, each with  4 cores and 8GB of  RAM. We  check memory
requirements:  173M x 4 x  4 = 2.77GB. This is  way below the  available RAM, so
there are no issues fitting the genome in memory. However, we  still want to use
both available machines, so we split the reads in two sets.

  $ grep '^>' reads.5m.50bp.cs.fa | wc -l
  5000000
  $ $SHRIMP_FOLDER/utils/splitreads.py 2500000 reads.5m.50bp.cs.fa
  $ ls *_to_*
  0_to_2499999.csfasta 2500000_to_4999999.csfasta
  $ grep '^>' 0_to_2499999.csfasta | wc -l
  2500000

This reads are now  in the two  files above. We  assume  the directory  with the
reads  as   well as the   directory  containing  SHRiMP2  are  shared across the
machines. We map the reads with the following commands

  <machine1>$ $SHRIMP_FOLDER/bin/gmapper-cs 0_to_2499999.csfasta \
  ciona.fa \
  -N 4 -o 5 -h 80% >map.0_to_2499999.out 2>map.0_to_2499999.log

  <machine2>$ $SHRIMP_FOLDER/bin/gmapper-cs 2500000_to_4999999.csfasta \
  ciona.fa \
  -N 4 -o 5 -h 80% >map.2500000_to_4999999.out 2>map.2500000_to_4999999.log

The options used above have the following meaning:
  -N 4 : Use 4 threads of control, corresponding to the number of cores
	 available on each machine.
  -o 5 : Print top 5 scoring hits for each read.

  -h 80% : Set the threshold score for a hit to 80% of the maximum possible
	   score. For reads of 50bp and the default match score of 10, the
	   maximum score is 500, and the threshold is set to 400.

Eventually, the two mapping jobs complete. We merge the mappings into one file
with:

  $ $SHRIMP_FOLDER/utils/merge-hits-diff-qr-same-db.py map.0_to_2499999.out \
  map.2500000_to_4999999.out >map.out

Above,  the merging script  we used   corresponds  to "different  query  set" (2
disjoint sets of reads), "same database" (the full ciona genome).


3.4 Mapping against a genome whose projection DOES NOT fit in RAM
-----------------------------------------------------------------

Assume we have:
./hg18.fa - fasta file containing all contigs of the reference human genome
./reads.500kx2.36bp.ls.fa - 1 million 36bp paired letter space reads

Assume we  know that the reads  in each pair come  from opposite  strands of the
genome,   pointing "inwards" (towards  each  other), with  an  insert of average
length 200 and standard deviation 10.

Assume we have a single machine with 2 quad-core CPUs and 16GB of RAM.

We look at the genome and the reads:

  $ ls -lh hg18.fa
  -rw-r--r-- 1 username 3.0G May  6 14:23 hg18.fa
  $ head -n 8 reads.500kx2.36bp.ls.fa
  >071214_EAS56_0061:1:1:882:575:1
  GAAATTGTACACATCACACACCTTTTTGTCTTTCTT
  >071214_EAS56_0061:1:1:882:575:2
  TGCTAGATAATGATTTGCACATTTTTAATTGGTAGA
  >071214_EAS56_0061:1:1:781:662:1
  GAAGTGATTACATTAAGCTGTACATATGCAGCATTC
  >071214_EAS56_0061:1:1:781:662:2
  AAATATGTTGATTTTTCAGTGTTTTTAAAGTAAAAT

We compute the memory requirements for mapping  against hg18 using the default 4
seeds: 3.0G x 4 x 4 =  48GB of RAM.  Our machine only has 16GB  of RAM. We split
the genome with the following utility:

  $ $SHRIMP_FOLDER/utils/split-db.py --ram-size 14 --prefix hg18 hg18.fa
  [...]
  $ ls *.fa
  hg18-14gb-12_12_12_12seeds-1of4.fa  hg18-14gb-12_12_12_12seeds-3of4.fa
  hg18-14gb-12_12_12_12seeds-2of4.fa  hg18-14gb-12_12_12_12seeds-4of4.fa

Above,  we decided to  give the gmapper process  only 14GB of RAM (the remaining
2GB are presumable used by the  operating system.) The  script split hg18 into 4
chunks, each containing several contigs, that will each fit in 14GB of RAM.

At this point, we realize we might be mapping letter space reads against hg18 in
the future, so we decide to create projections  of the individual genome pieces,
so that we avoid having to re-project them every time we start a mapping job. To
achieve this, we run:

  $ $SHRIMP_FOLDER/utils/project-db.py --shrimp-mode ls hg18-14gb-*.fa
  [...]
  $ ls hg18-14gb-*-ls.*
  hg18-14gb-12_12_12_12seeds-1of4-ls.genome
  hg18-14gb-12_12_12_12seeds-1of4-ls.seed.0
  hg18-14gb-12_12_12_12seeds-1of4-ls.seed.1
  hg18-14gb-12_12_12_12seeds-1of4-ls.seed.2
  hg18-14gb-12_12_12_12seeds-1of4-ls.seed.3
  hg18-14gb-12_12_12_12seeds-2of4-ls.genome
  hg18-14gb-12_12_12_12seeds-2of4-ls.seed.0
  hg18-14gb-12_12_12_12seeds-2of4-ls.seed.1
  hg18-14gb-12_12_12_12seeds-2of4-ls.seed.2
  hg18-14gb-12_12_12_12seeds-2of4-ls.seed.3
  hg18-14gb-12_12_12_12seeds-3of4-ls.genome
  hg18-14gb-12_12_12_12seeds-3of4-ls.seed.0
  hg18-14gb-12_12_12_12seeds-3of4-ls.seed.1
  hg18-14gb-12_12_12_12seeds-3of4-ls.seed.2
  hg18-14gb-12_12_12_12seeds-3of4-ls.seed.3
  hg18-14gb-12_12_12_12seeds-4of4-ls.genome
  hg18-14gb-12_12_12_12seeds-4of4-ls.seed.0
  hg18-14gb-12_12_12_12seeds-4of4-ls.seed.1
  hg18-14gb-12_12_12_12seeds-4of4-ls.seed.2
  hg18-14gb-12_12_12_12seeds-4of4-ls.seed.3

Since we have a single machine, we run the mapping process sequentially against
each of the chunks.

  $ for i in 1 2 3 4; do \
      $SHRIMP_FOLDER/bin/gmapper-ls -L hg18-14gb-12_12_12_12seeds-${i}of4-ls \
      reads.500kx2.36bp.ls.fa \
      -N 8 -p opp-in -I 50,500 -m 20 -i -25 -g -40 -e -10 -E \
      >map.db${i}of4.sam 2>map.db${i}of4.log
    done
  [...]
  $ ls map.db*.sam
  map.db1of4.sam map.db2of4.sam map.db3of4.sam map.db4of4.sam

The options used above have the following meaning:
  -L hg18-14gb-12_12_12_12seeds-${i}of4-ls : Load the projection of the i-th
    chunk of the genome. This is in the 5 files with this prefix, and ending in
    .genome and .seed[1-4]
  -N 8 : Use 8 threads of control, fully utilizing 2 quad-core CPUs.
  -p opp-in : Enabled paired mode, where the two reads in each pair are on
    "opp"-osite strands, pointing "in"-wards.
  -I 50,500 : Set minimum and maximum allowed insert size.
  -m 20 -i -25 -g -40 -e -10 : Set specific Smith-Waterman scores for match,
    mismatch, gap open and gap extend.
  -E : Enable SAM output.

After the mapping jobs complete, we merge the results with:

  $ $SHRIMP_FOLDER/bin/mergesam reads.500kx2.36bp.ls.fa map.db?of4.sam > map.sam

Above, the merging script we used corresponds to "same query set" (same reads)
and "different databases" (disjoint chunks of hg18).

NOTE: For an example in which we split both the reads and the genome, combining
examples 3.3 and 3.4, see $SHRIMP_FOLDER/SPLITTING_AND_MERGING.


3.5 Mapping cDNA reads against a miRNA database
-----------------------------------------------

Assume we have:
./hsa.miRNA.cDNA.fa - database of mature miRNA sequences in human (as cDNA)
./reads.1m.cDNA.50bp.csfasta - 1 million 50bp colour space cDNA reads

We project the database with:

  $ $SHRIMP_FOLDER/utils/project-db.py --seed 00111111001111111100,\
  00111111110011111100,00111111111100111100,00111111111111001100,\
  00111111111111110000 --h-flag --shrimp-mode cs hsa.miRNA.cDNA.fa
  $ ls hsa.miRNA.cDNA-ls.*
  hsa.miRNA.cDNA-ls.genome  hsa.miRNA.cDNA-ls.seed.1  hsa.miRNA.cDNA-ls.seed.3
  hsa.miRNA.cDNA-ls.seed.0  hsa.miRNA.cDNA-ls.seed.2  hsa.miRNA.cDNA-ls.seed.4

Here, we  used 5  seeds of weight   14 and span   20 which  allow for  2 initial
mismatches (the leading 00), followed by a match of length  7bp (the 6 following
1s), followed by another match of various forms.

Note:  The seeds  are  applied to  the colour   space  translations of both  the
database and the  reads. Hence, a  string of 6  consecutive 1s corresponds  to 6
consecutive matching colours, which are in particular generated by a string of 7
consecutive matching letters.

We map the reads with:

  $ $SHRIMP_FOLDER/bin/gmapper-cs -L hsa.miRNA.cDNA-ls \
  reads.1m.cDNA.50bp.csfasta \
  -N 8 -n 1 -U -o 2 -F -v 50% -h 50% -P \
  >map.out 2>map.log

The options used above have the following meaning:
  -L hsa.miRNA.cDNA-ls : Load genome projection.
  -N 8 : Use 8 threads of control.
  -n 1 : Require a single space seed match to investigate region.
  -U : Perform ungapped alignment.
  -o 2 : Print 2 top hits per read.
  -F : Process only the forward strand.
  -v 50% -h 50% : Set the score thresholds for the SW filtering steps to 50% of
    the maximum. This effectively disables the filtering by score threshold.
  -P :  Pretty-print mappings so that we can look at them.

To clarify the settings above:
- only the forward strand (-F) of database sequences is inspected;
- mapping regions are selected by a single spaced seed match (-n 1);
- ungapped (-U) mapping scores are still being computed (only match, mismatch
and crossover scores are now relevant);
- mappings that do not achieve a score of 50% of the maximum are discarded (not
many of those, because the initial seed match already guarantees a decent
score);
- the 2 top hits (-o 2) per read are output.

In SHRiMP 2.0.4, a new command-line option was introduced that can be used to
load several miRNA-specific settings: "-M mirna" or "--mode mirna" is
equivalent with:
  loading the 5 seeds mentioned above; plus
  "-H -n 1 -w 100% -U -a 0 -g -255 -q -255 -Z"

We have a look at the mappings:

  $ head -n 15 map.out
  #FORMAT: readname contigname strand contigstart contigend readstart readend r\
  eadlength score editstring
  >1380_607_170_F3        hsa-let-7e MIMAT0000066 +       1       20      1    \
     20      35      186     15x5

  G:          1    TGAGGTAGGAGGTTGTATAGTT-------------    20
                   |||||||||||||||X||||
  T:               TGAGGTAGGAGGTTGtATAG---------------
  R:          1   T01220132022010123332210233322123032    20

  >1380_721_321_F3        hsa-miR-524-5p MIMAT0002849     +       1       22   \
     1       22      35      220     22

  G:          1    CTACAAAGGGAAGCACTTTCTC-------------    22
                   ||||||||||||||||||||||
  T:               CTACAAAGGGAAGCACTTTCTC-------------
  R:          1   T22311002002023112002222302010303102    22



3.6 Advanced: Mapping in OVERLY sensitive mode
----------------------------------------------

SHRiMP2 is  designed to be quite   sensitive, but without  complete disregard to
speed.  The following  is an ADVANCED  EXAMPLE showing  some  settings that will
make it OVERLY sensitive.

CAUTION: Not all of these settings need to be  applied at the  same time to make
gmapper  sensitive   enough for   your needs.     Please read  this  example  in
conjunction with the relevant  descriptions in the  Algorithms Section, and make
an informed choice of parameters for your application.

Assume we have  a set  of (not  so many...) reads   that we really need  to find
mappings for, and the appropriate genome projection. We can try:

  $ $SHRIMP_FOLDER/bin/gmapper-cs -L database reads.csfasta \
  -V -w 150% -n 1 -r 50% -v 55% -l 40% -Z -h 60% -a -1 \
  >map.out 2>map.log

The significance of these options  is  the following:

  -V : Do not automatically  trim genome index list  that are unusually long. In
  our  tests with hg18,  this results in about  1-2% more hits,  but the running
  time can increase dramatically, to 3x or more.

  -w 150% : Enlarge the length  of the genome window against  which each read is
  being mapped.

  -n  1 : Enable  window  generation mapping mode  1.   This is very costly when
  working  with a large genome.  The  setting means  a  single spaced kmer match
  between a  read and the  genome can  be  enough to create a  candidate mapping
  window. With  seeds of weight  12 and a  uniform random genome, a random match
  occurs once  in every 4^12   locations. For 1/4 of hg18,   this means about 44
  random matches, for  every spaced kmer  and  for every seed.  With  4 seeds of
  average span 20 and  reads  of length  50, thats  about (50-20)x4x44 ~=  5,000
  random single spaced kmer  matches. Investigating a candidate window  location
  around each of  those is costly.   The default  window generation  mode (-n 2)
  requires at least 2 spaced kmer matches, which reduces dramatically the number
  of  windows created around  random matches. (Even  though a match  of 2 spaced
  seeds of weight 12 in a  genome window does not  guarantee 2x12=24 matches, by
  design  of the  seeds it  does guarantee  about  17 matches. As  a result, the
  number  of windows expected  to  be placed around  random  matches drops by  a
  factor of at least 4^5 = 1024.)

  For paired mapping mode, you can try the window generation  mode "-n 3" first,
  and, at the extreme, "-n 2".

  -r 50%  :  Set    the window  generation    threshold  to 50%  of   the   read
  length.  Combined with -n 1,  the effect of this  is  that every single spaced
  kmer match from a seed with span greater than 50% of read length will generate
  a candidate mapping window.

  -v 55% : Lower the threshold for the vector SW filter.

  -l 40% :   Increase the  allowable  overlap of   one mapping location   with a
  previously  inspected mapping  location that passed   the vector filter.  This
  helps if for some reason the  data makes it hard for  gmapper to center genome
  mapping locations around spaced kmer matches.

  -Z : Disable caching of previous vector SW runs.

  -h 60% : Lower the threshold for the full SW filter.

  -a -1 :  Disable the restriction of  the full SW dynamic programming algorithm
  to areas  around the  matching spaced kmers.   This  can help detecting larger
  indels,  but the  runtime of the  full  SW  filter  will increase drastically.
  Still, in some cases this is  ok to do, as  this runtime is much smaller than,
  say, the time spent in the vector SW filter. See the log  file from a relevant
  run for these timings.


--------------------------------------------------------------------------------
4. Input and Output Specifications
--------------------------------------------------------------------------------

Input Files
-----------

  Both the reads and the  genome files should be in  fasta format. The files are
  opened using the  zlib library, so  it is ok for them  to be gzip-ed. (But not
  zip-ed!)

  The  genome  fasta should   _always_     be  in  letter  space.   The    space
  (letter/colour) of the reads should match the  gmapper mode, which is selected
  by invoking gmapper by the appropriate symlink, gmapper-ls or gmapper-cs.


Output Files
------------

  SHRiMP2 supports SAM output,   as well as  native  SHRiMP output.  SAM  output
  format  is  documented elsewhere.   SHRiMP   output  format  is  remains  from
  SHRiMP1. For more details, see Section 7.

  In either  output format (except  for SHRiMP with  pretty-print), each mapping
  occupies one line. The top <num_outputs> mappings for  each read appear in the
  output file in the order of best to worst.

  When running gmapper in multi-threaded mode, mappings of different reads might
  appear in  the output file intermingled, as  follows (only displaying the name
  and the score):

    $ grep "^>" map.out | head -n 8
    >read_1 [...] 355
    >read_1 [...] 355
    >read_2 [...] 376
    >read_2 [...] 376
    >read_1 [...] 327
    >read_3 [...] 415
    >read_1 [...] 327
    >read_2 [...] 376

  Several scripts are provided that manipulate output files (merge, sort).


Paired Mapping Mode Input
-------------------------

  When  running in paired   mode,  the two   reads in each  pair  should  appear
  consecutively in the reads file, as follows:

    $ head -n 8 reads.500kx2.36bp.ls.fa
    >071214_EAS56_0061:1:1:882:575:1
    GAAATTGTACACATCACACACCTTTTTGTCTTTCTT
    >071214_EAS56_0061:1:1:882:575:2
    TGCTAGATAATGATTTGCACATTTTTAATTGGTAGA
    >071214_EAS56_0061:1:1:781:662:1
    GAAGTGATTACATTAAGCTGTACATATGCAGCATTC
    >071214_EAS56_0061:1:1:781:662:2
    AAATATGTTGATTTTTCAGTGTTTTTAAAGTAAAAT

  In particular:
    1. The reads should not be simply dumped in the same file. That is, the
    reads in each pair should not be just about anywhere in the file. They must
    be consecutive.
    2. There should not be unpaired reads among the paired reads. If you have
    both paired and unpaired reads, use different files and different runs.

  Several scripts are provided to get the reads in this format.

  Alternatively,  the  reads in  each file  can be  placed  in two corresponding
  files, which are then loaded using options -1 and -2.

  CAUTION: gmapper will NOT check any of the above. If started in paired mode,
  it will simply fetch consecutive reads and assume they form a pair.

  For SAM output, gmapper infers the name of a pair of reads as the longest
  common prefix of the two read names in that pair. E.g., the first pair above
  would be called "071214_EAS56_0061:1:1:882:575:", including the trailing ":".


Paired Mapping Mode Output
--------------------------

  In paired mapping  mode, gmapper outputs  pairs of mappings,  which consist of
  consecutive lines describing   the  mapping  of    the  first read    followed
  immediately by the mapping of the second read. Even though several mappings of
  the  same pair need not occur  consecutively (because of multi-threading), the
  two mappings inside each pair _do_ occur consecutively. For example,

    >071214_EAS56_0061:1:1:882:575:1 [...] 350
    >071214_EAS56_0061:1:1:882:575:2 [...] 310
    >071214_EAS56_0061:1:1:781:662:1 [...] 320
    >071214_EAS56_0061:1:1:781:662:2 [...] 360
    >071214_EAS56_0061:1:1:781:662:1 [...] 320
    >071214_EAS56_0061:1:1:781:662:2 [...] 340
    >071214_EAS56_0061:1:1:882:575:1 [...] 310
    >071214_EAS56_0061:1:1:882:575:2 [...] 350

  Scripts are provided that sort mappings of pairs.


--------------------------------------------------------------------------------
5. Parameters
--------------------------------------------------------------------------------

The following parameters govern the  behaviour of  gmapper.   For more  thorough
descriptions of the algorithms involved, see the next Section.

The list is organized conceptually by the "area" where parameters operate.


Genome Projection
-----------------

  [ -s/--seeds <seed1,seed2,....> ]

    Each spaced seed is a  contiguous string of 0s  and 1s. Several spaced seeds
    can be specified either  by separating them by commas  as a single parameter
    to -s, or  by giving -s several times.  By default, gmapper uses  a set of 4
    spaced seeds of weight 12. Several sets  of default seeds are available  for
    certain weights. These can be loaded by, e.g. "-s w16",  to load the default
    set of seeds of weight 16.

    Note: When running gmapper in colour space mode, seeds are applied to colour
    space representations of the genome and the reads.

  [ -H/--hash-spaced-kmers ]

    Hash spaced kmers obtained from each spaced  seed into 24-bit strings before
    indexing them.

  [ -z/--cutoff <cutoff> ]

    Ignore lists in the genome index that are longer than <cutoff>.

  [ -V/--trim-off ]

    Disable automatic genome  index trimming. By default,  if no  "-z" is given,
    gmapper trims the lists in the  genome index longer  than a given threshold.
    Currently, the automatic threshold is (about):

      max(1000, 100*<average_list_length>).

    This flag is ignored if either -z or -S is given.

  [ -S/--save <filename> ]

    With this parameter,  gmapper projects and indexes  the genome, and saves it
    in several  files for  future use.  No read mapping   is performed. The only
    other relevant parameters in this mode of operation are: -s, -H, and -z.

    The files created are:

    <filename>.genome
      - contains the genome sequence, along with some extra data

    <filename>.seed.0, <filename>.seed.1, ...
      - each such file contains the index of  the genome projected by one spaced
      seed.

    Note: If using -S and -L to save  and load a genome  index, note that letter
    and  colour space indexes are different.  If you have  colour (letter) space
    reads, you should build a colour (letter) space index of the genome.

  [ -L/--load <filename> ]

    Load a genome index from the given file. Also  loads the set of spaced seeds
    used in the projection, as  well as the value of  the -H flag when the index
    was created. Any -s or -H flags are ignored.

    In  "short form", <filename>  does not contain the  character ",", and it is
    interpreted as the prefix  of the path to the  index files. The actual files
    are then found by appending ".genome" and ".seed.*".

    In "long  form", <filename> contains  the character  ",".  In this case, the
    string is tokenized by the "," separator, obtaining, say, "A,B,C". Then, "A"
    is interpreted as the  .genome file, and "B" and  "C" are interpreted as the
    .seed.* files.

    Thus, if a genome index was created using 2 seeds and the parameter "-S db",
    then  "-L  db"   is equivalent to    "-L db.genome,db.seed.0,db.seed.1". The
    advantage of the long form of -L is that one may  specify which of the seeds
    to load, rather than loading them all, which is  what happens with the short
    form. E.g., we could use "-L db.genome,db.seed.1".

  [    --save-mmap /<mmap_name> ]
  [    --load-mmap /<mmap_name> ]

    Can be used to save and subsequently load a genome projection  to  and  from
    shared memory.  The projection will remain resident in shared  memory  until
    explicitly removed with

    $ rm /dev/shm/<mmap_name>

    When saving to shared memory,  the genome projection must  be  first  loaded
    with -L (not directly from a fasta file). The name parameter must start with
    a '/' character, and it must contain no other '/' characters.

    This functionality is useful only when  many  individual  gmapper  runs  are
    performed against the same large genome projection,  to the point where  the
    projection loading part (with -L) is a bottleneck.

    For reasons we did not fully investigate,  loading  does not work on certain
    machines. Consequently, the functionality should be considered experimental.


General Mapping
---------------

  [ -w/--match-window <window_length> ]

    Specifies the length of the genome window against which a read is mapped. It
    can be given either as an absolute value, e.g., "-w  62", or as a percentage
    of read length, e.g., "-w 133%". It defaults to "-w 140%".

  [ -U/--ungapped ]

    Perform ungapped alignment.

  [ -F/--positive ]

    Only process the forward (positive) strand of the genome.

  [ -C/--negative ]

    Only process the reverse complement (negative) strand of the genome.


Smith-Waterman Scores
---------------------

  [ -m/--match <match_score> ]

    The score for matches during the Smith-Waterman score calculation. This
    should be positive.

  [ -i/--mismatch <mismatch_score> ]

    The score for mismatches. This should be negative.

  [ -g/--open-r <gap_open_reference_score> ]

    The score to open a gap along the genome sequence. Should be negative.

    Note: In the current implementation, the gap_open score does not include any
    extension.  That is, a gap of length 1 is scored as:

      <gap_open_score>+<gap_extend_score>

  [ -q/--open-q <gap_open_query_score> ]

    The score to open a gap along the read sequence. Should be negative.

    Note: If -g is  set and -q is  not set, the  gap open penalty for the  query
    will be set to the same value as specified for the reference.

  [ -e/--ext-r <gap_extend_reference_score> ]

    The score to extend a gap along the genome sequence. Should be negative.

  [ -f/--ext-q <gap_extend_query_score> ]

    The score to extend a gap along the read sequence. Should be negative.

    Note: If -e is set and -f  is not set, the  gap extend penalty for the query
    will be set to the same value as specified for the reference.

  [ -x/--crossover <crossover_score> ]

    The score for  calling a colour space sequencing  error. Should be negative.
    Available only in colour  space mode. Disable  with a prohibitive value: "-x
    -255".

  [    --pr-xover <pr_xover> ]

    Sets the default crossover  probability.  This  is  used  to  compute  other
    probabilities from scores.  During  alignment,  the  probability a colour is
    wrong is derived from its quality values  (if available)  rather than  using
    this default. (Default: 0.03)


Filter 1: Window Generation
---------------------------

  [ -n/--cmw-mode <window_generation_mode> ]

    This parameter controls how candidate mapping windows  (CMW) are created for
    every  read or  readpair.  In   unpaired mode,  possible values are   1 or 2
    (=default). In paired more, possible values are 2, 3, or 4 (=default). Their
    significance is as follows.

    In unpaired mode, create a CMW for a read based on:
      1: as little as 1 spaced seed match
      2: at least 2 spaced seed matches

    In paired mode, create 2 CMWs, one for each foot (read in a pair), based on:
      2: as little as 1 spaced seed match in either foot
      3: at least  2 spaced seed matches in one foot,  and as little as 1 in the
	 other
      4: at least 2 spaced seed matches in either foot

  [ -r/--cmw-threshold <window_generation_threshold> ]

    Generate a CMW as long as the "optimistic" estimation of  the match score is
    greater than  <window_generation_threshold>. The estimation  is computed  by
    assuming  that all  positions inside  a   spaced kmer  match  are, in  fact,
    matching. The threshold is either absolute, e.g., "-r  252", or a percentage
    of  the maximum possible  match score,  e.g.,  "-r 50%". It  defaults to "-r
    55%".

    Note: This threshold is ignored  when  either <window_generation_mode> is  1
    (in unpaired mode  this happens with "-n  1"; in paired  mode with "-n 2" or
    "-n 3"), or when performing ungapped alignment ("-U").

    Note on 454 reads
    -----------------

    Currently, gmapper constructs the estimation is this step based on at most a
    single indel. This is appropriate for read lengths less than 100. For longer
    reads,  which might contain  more indels,  you might  want to  significantly
    lower this threshold, and use an absolute value. For example, to allow a CMW
    to pass this filter based on two spaced kmer matches  of total length 50 and
    one indel (with the default  scores), you should use  "-r 464". (We get  464
    from 50 matches: +500, 1 gap open: -33, 1 ins extend: -3.)


Filter 2: Vector SW/Ungapped Alignment
--------------------------------------

  [ -v/--vec-threshold <vector_sw_threshold> ]

    For every  CMW  for a  read,  discard  it if  after  running the  vector  SW
    algorithm, its mapping  score is below  <vector_sw_threshold>. The threshold
    is either absolute, e.g., "-v 300", or a  percentage of the maximum possible
    match score, e.g., "-v 60%". It defaults to "-v 60%".

    Available only in   colour space mode.  (In  letter space mode,  this equals
    <full_sw_threshold>.)

  [ -l/--cmw-overlap <window_overlap> ]

    Following a CMW filter 2  hit, ignore further  CMWs that overlap it by  more
    than <window_overlap>. It  can be given as  either an absolute value,  e.g.,
    "-w 62", or a percentage of read length, e.g.,  "-w 50%". It defaults to "-l
    80%".

  [ -Z/--cachebypass-off ]

    Disable cache  bypass of vector SW calls.   This  is disabled by  default in
    ungapped mode ("-U").


Filter 3: Scalar (Full) SW Alignment
------------------------------------

  [ -h/--full-threshold <full_sw_threshold> ]

    For  every  CMW, discard  it  if after  running  the  full SW algorithm, its
    mapping   score is   below  <full_sw_threshold>.  The  threshold is   either
    absolute,  e.g., "-h  400", or a  percentage of  the  maximum possible match
    score, e.g., "-h 80%". It defaults to "-h 68%".

  [ -o/--report <num_outputs> ]

    Output the top <num_outputs> filter 3 hits for each read/pair.

  [    --max-alignments <n> ]

    If more  than <n>  mappings for a  read (pair)  pass all filters,  drop them
    all. Also see Note on Post-alignment Option Ordering.

  [    --strata ]

    Only  output  the  highest  scoring  mappings  for any  given  read,  up  to
    <num_outputs>  mappings per read.   Also see  Note on  Post-alignment Option
    Ordering.

  [ -a/--anchor-width <anchor_width> ]

    Limit the full SW  algorithm to a  band  of <anchor_width> additional  width
    around matching kmers (anchors). This defaults to "-a 8" in gapped mode, and
    "-a 0" in ungapped mode. To disable anchor usage altogether, use "-a -1".

  [ -T/--rev-tiebreak ]

    Reverse the order in which tie-breaks are resolved on the negative strand.

  [ -t/--tiebreak-off ]

    Disable reverse tiebreak (-T/--rev-tiebreak , on by default).

  [    --global ]

    Perform a full global alignment in the last phase of alignments.

    NOTE: This is on by default as of v2.2.0.

  [    --local ]

    Perform local alignment instead of global. In this mode, the ends of a read
    do not necessarily need to match the reference.  Mapping qualities are  not
    available in this mode.

    NOTE: This used to be the default setting prior to v2.2.0.

  [    --bfast ]

    Color-space only. Try to align like BFAST. This enables global alignments
    and if in FASTQ mode, outputs base qualities in the QUAL field, just like 
    BFAST. 

  [    --indel-taboo-len <value> ]

    Prevent indels from starting or ending in the tail <value> positions of a
    read. Note: for deletions, start and end positions are the same with
    respect to the read. Insertions might still start before the taboo zone,
    but then they have to go all the way to the end of the read (this will
    only ever happen with --global). Only available in colour space.


Sets of settings
----------------

  [ -M/--mode <mode> ]

    Currently only "mirna" is a valid mode. In this case, the option is
    equivalent with:
      loading the 5 seeds mentioned above (in the miRNA section 3.5); plus
      "-H -n 1 -w 100% -U -a 0 -g -255 -q -255 -Z"


Paired Mode
-----------

  [ -p/--pair-mode <paired_mode> ]

    By  default,   gmapper operates in unpaired   mapping  mode.  This parameter
    enables  the paired mapping     mode.  The possible  values are:   "opp-in",
    "opp-out", "col-fw" and  "col-bw".  These correspond  to: opposing  strands,
    inwards; opposing strands,  outwards; colinear, second is forward; colinear,
    second is backward.

  To illustrate the settings, consider the following genome and perfect reads:

         -->   -->
         GGA   GTA
      + TGGATCCGTAA
      - ACCTAGGCATT
         CCT   CAT
         <--   <--

    (R1:GGA, R2:TAC) or (R1:TAC, R2:GGA) are "opp-in", insert size = 8
    (R1:GTA, R2:TCC) or (R1:TCC, R2:GTA) are "opp-out", insert size = 4
    (R1:GGA, R2:GTA) or (R1:TAC, R2:TCC) are "col-fw", insert size = 6
    (R1:GTA, R2:GGA) or (R1:TCC, R2:TAC) are "col-bw", insert size = 6

  [ -I/--isize <min,max> ]

    Limit the search  space  for paired mappings to those  where the insert size
    between the two feet falls in  the  given  range.  This  parameter  is  only
    available in paired mapping mode. It defaults to "-I 0,1000".

    In all paired  mapping modes, the insert is defined  as the distance between
    the 5' ends  of the reads (see  example above.) Neither the min  nor the max
    values can be negative. If you are unsure about the orientation of the input
    reads, try each pairing mode along with -X on a statistically non-negligible
    set of reads (say, 10,000). The  correct pairing mode will likely be the one
    where the  insert size histogram is  neither flat, nor  abundant in negative
    values.

    NOTE:  This limit is NOT strictly enforced.  By  nature  of  the  underlying
    algorithms,  some paired mappings with inserts just outside the given  range
    will also be considered.  These are not  discarded  by  gmapper  itself.  To
    strictly enforce these bounds, use a script to filter the SAM output.

  [    --half-paired ]

    In paired mode, if a pair does not map, try to map each read individually.
    (This is on by default as of v2.2.0).
 
  [    --no-half-paired ]

    Do NOT try to map individual reads.  If a pair does not map,  do NOT try  to
    align each read independently. (The default is to try.)

  [    --sam-r2 ]
    Report the SAM r2 field for letter space alignments. Report a similar x2 fi-
    eld for colour space alignments.

  [    --insert-size-dist <mean,stddev> ]

    Specifies the mean and standard deviation of  the  insert  sizes.  They  are
    assumed to come from a normal distribution. These values are used in mapping
    quality computation. The defaults are: mean=200, stddev=100.


Thread Control
--------------

  [ -N/--threads <num_threads> ]

    Use <num_threads> threads of control. It defaults to "-N 1".

  [ -K/--thread-chunk <chunk_size> ]

    Threads take turns  "checking out" a  chunk of reads, mapping them, printing
    the results, and "checking out" the next chunk. This parameter specifies how
    many reads should be in each such chunk. Defaults to "-K 1000".

  [ -D/--thread-stats ]

    Print individual thread statistics in the log file.


Input Control
--------------

  [ --longest-read ]

    The longest input read is of this length.

  [ -Q/--fastq ]

    The input is FASTQ format.

  [ -1/--upstream <filename> ]

    Use the given file as input for the upstream reads in paired mode.

  [ -2/--downstream <filename> ]

    Use the given file as input for the downstream reads in paired mode.

  [    --min-avg-qv <value> ]

    The minimum average quality value of a read for it to even be considered for
    mapping. The default is 10.

  [    --qv-offset <value> ]

    Interpret qvs in fastq input as PHRED+<value>.  The default is 33 for colour
    space and 64 for letter space.

  [    --no-qv-check ]

    By default, gmapper crashes if it detects a qv less than -10 or greater than
    50.  This  helps  prevent mix-up between various input formats.  This option
    diables this check.

  [    --ignore-qvs ]

    When input is fastq, completely ignore base/colour qvs from the analysis, as
    if the input were fasta.

  [    --trim-front <value> ]
  
    Trim the front of read (or both reads of pair), by this amount.

  [    --trim-end <value> ]
  
    Trim the end of read (or both reads of pair), by this amount.

  [    --trim-first ]
  
    Trim only the first read in read pair.

  [    --trim-second ]

    Trim only the second read in the read pair.

  [    --trim-illumina ]

    Trim trailing sequence that has 'B' (2) quality values. 


Output Control
--------------

  [    --shrimp-format ]

    Select old SHRiMP output format.

  [ -P/--pretty ]

    Select SHRiMP output format, and pretty-print alignments.

  [ -R/--print-reads ]

    Select SHRiMP output format, and include reads in the output.

  [ -E/--sam ]

    Select SAM output format. (This is on by default as of v2.2.0.)

  [ --sam-unaligned ]

    If SAM  output format is  also selected  dump   unaligned reads to  the  SAM
    output.

  [ --sam-header <filename> ]

    Output   the  filename  as header    instead    of regular   SHRiMP  /   SAM
    header. Includes additional lines defining a read group if need be.

  [ --sam-header-XX <filename> ]

    Where XX is one of "hd", "sq", "rg", "pg". Replace the normal gmapper output
    for that line with the contents of that file.

    Note: If --sam-header is given, all headers are taken from  that file and no
    additional ones  are printed.  If some --sam-header-XX   are given  and some
    aren't, the given  ones  are taken  from the associated  files, and  the not
    given ones are filled in by gmapper.
  
  [ --read-group read_group,pool_name ]

    Include in the SAM  output the corresponding  read_group and pool_name. This
    is ignored if either --sam-header or --sam-header-rg are given.

  [ --un <filename> ]
  [ --al <filename> ]

    Print unaligned/aligned reads to target file in the same format as the input
    (fasta or fastq).

    NOTE: Currently, the  reads in these files  are not guaranteed to  be sorted
    in the same order as in the input when multiple threads are used!

  [    --progress <value> ]

    Display a progress line each <value> reads. the default is 100,000.


Mapping Qualities
-----------------

  [    --no-mapping-qualities ]

    Do not compute mapping qualities. By default, they are computed.

  [    --all-contigs ]

    This flag specifies  that all contigs of interest are  included in this run.
    When this  is not the  case (which is  the default), gmapper  includes extra
    fields in  its SAM  output in  order to make  possible the  recalculation of
    mapping qualities when merging mappings from different files.  Also see Note
    on Post-alignment Option Ordering.

  [    --single-best-mapping ]

    If mapping qualities are computed,  this flag tells gmapper to  only  output
    the "best" mapping for each read (or pair, in paired mode).

    Note:   in  paired  mode,   if  --single-best-mapping   is  given   but  not
    --all-contigs,  gmapper  outputs the  best  mapping  from  each class:  best
    paired, best unpaired  for the first read, and best  unpaired for the second
    read.  These extra  mappings are  used to  recompute mapping  qualities when
    merging.  Also see Note on Post-alignment Option Ordering.

  [    --no-improper-mappings ]

    Only relevant in paired  mode, when --all-contigs  and --single-best-mapping
    are both given. By default, if SHRiMP decides the best  mapping of a pair is
    a singleton,  it will attempt  to pair it  up with  the best mapping  of the
    other read,  if the latter  is  strong enough. This   enables mapping in the
    presence of large structural variations. The flag disables this behaviour.


Diagnostics
-----------

  [ -X/--isize-histogram ]

    When running in paired mapping mode, dump histogram of insert sizes. (Use it
    with a statistically significant small  set of reads.)  This can be used  to
    check the correct  paired mapping  mode was  selected. The distribution   of
    insert sizes should be bell-shaped. If it looks uniform, the paired mode was
    probably wrong.

  [ -Y/--proj-histogram ]

    Dump histogram of  the lengths of  the genomic index lists. The distribution
    of these list lengths has huge tails (e.g. average 50, maximum 500,000), and
    trimming some of the outliers  often results in  a  dramatic speed boost  at
    very little, if any, cost to sensitivity.


Note on Post-alignment Option Ordering
--------------------------------------

  The set of mappings which pass the Scalar/Full SW threshold are then processed
  in this order:

    (1) If --strata  is given, it is applied  first, based on (sum of) scores.

    (2) If --max-alignments is gien, it is applied second.

    (3) Mapping qualities are then computed based on the remaining reads.

    (4) If --single-best-mapping is given for paired reads:

      (4a) If --all-contigs is not given,  at most one top mapping in each class
      is reported, where  the possible classes are: singleton  mapping for first
      read, singleton mapping for second  read, paired mapping. Here, sorting is
      by mapping quality.

      (4b) If  --all-contigs is given,  at most one  mapping is output  for each
      read pair. Selection between classes is done based on mapping quality.

      Note: For unpaired reads, there is  only one class of mapping, so (4a) and
      (4b) are equivalent  in terms of what mappings  are output.  However, they
      are different  in that --all-contigs  also prevents additional  SAM fields
      from being included in the output which would have been used by merging.

--------------------------------------------------------------------------------
6. Algorithms
--------------------------------------------------------------------------------

The mapping algorithm (gmapper) works as follows. First, it projects and indexes
the genome using every  spaced seed. These indexes  are  all kept in  RAM. Next,
several  threads of control  are created.  Each  thread "checks out"  a chunk of
reads to be   mapped, it goes on   running independently of the   other threads,
mapping the  reads in its  chunk, it prints the mappings  as they are found, and
when it is done, it returns to "check out" the next chunk.

To map every read Q,  the set of all  possible mapping locations is conceptually
passed through 3 filters:  Window Generation, Vector SW/Ungapped Alignment,  and
Scalar (Full) SW Alignment. At the end, the top <num_outputs> remaining hits for
every read are output.

When running in paired mode, gmapper only looks  for locations where to map each
pair of reads in such a way that the size of the insert  between the reads falls
within a  specified range. In most cases,  this dramatically reduces the size of
the search space.

More detailed descriptions follow.


Genome Projection
-----------------

The first step in a gmapper  run is to project and  index the genome using every
spaced  seed. For a spaced  seed S of weight  (=number of 1s)  W,  the number of
possible spaced  kmers we might obtain  by applying S  to a genome  or a read is
4^W. (We drop kmers containing wildcard bases such as "N".) Thus, for every seed
S we build  an array of size 4^W,  where the location indexed  by R contains the
list of genome locations X, such that applying S at X yields the spaced kmer R.

The size of the spaced kmer indexes (4^W) quickly  becomes impractical. For this
reason, we offer the option (using the -H flag) to first  hash spaced kmers into
24-bit strings, then  store genome locations indexed by  those hash values. This
is equivalent  to  topping  the  spaced kmer index   size  at 2^24  =  4^12. The
disadvantage of this  scheme is that, during  the  matching, some of the  genome
locations that are investigated do not in fact contain matches to the read being
processed.


Trimming the Genome Index
-------------------------

In our experience,  the distribution of the  lengths  of the  lists in a  genome
index has huge    tails. The super-long lists  negatively   impact running time,
without providing any significant sensitivity gains.  E.g., for pieces of 1/4 of
hg18, the average list length is about 50, and the longest list has length about
500,000. Trimming (discarding) genome  lists longer than  about 5,000 lowers the
running time by a factor of about 3x, and decreases the number of hits output by
merely 1-2%.

Because of this dramatic benefit, gmapper currently implements a basic automatic
trimming step, selecting the cutoff value of:

  max( 1000, 100*4^total_genome_length / 4^max_seed_weight )

Informally,  this is either 1000  or 100x the  average list length, whichever is
higher.

The choice  described above is  very basic, and unfortunately, we  do not have a
reliable method to do  a better automatic  trimming. (This is  high on our to-do
list for the next release.) In our experience,  finding a good list_cutoff value
can be done as follows:

  - Split and project the genome untrimmed.

  - Run gmapper with the -Y -V  parameters  to see a list length histogram with:
    $ gmapper-ls -L genome -Y -V /dev/null >/dev/null

  - Based on histogram above, find a candidate value for list_cutoff, e.g.,
  8000;

  - Run  gmapper  with  several  list_cutoff values on  a  small  (say, 10k) but
  statistically significant subset of reads; discard the mappings, but watch the
  running times and the percentage of reads mapped in the log file. E.g.,

    $ for Z in 1000 2000 4000 8000 16000 32000 64000; do
        gmapper-ls -L genome reads.lsfasta -N 8 -z $Z >/dev/null 2>map.z${Z}.log
      done;
    $ grep -E "(Cutoff)|(Run-time:)|(Matched:)|(Matches:)" map.z*.log

  - After deciding on a  good value, you can either  give it on the command line
  for  every mapping job with  the -z parameter,  or you can construct a trimmed
  genome index with, e.g.:

    $ gmapper-ls -L genome -S genome-trimmed -z 4000


Filter 1: Window Generation
---------------------------

In the first filter, a read is projected using all spaced seeds, obtaining a set
of several spaced   kmers. For every  spaced seed   S  and every spaced   kmer R
obtained from Q using S, the corresponding genome index is looked up, retrieving
the list of genome locations where  we also obtain R  by applying S. Visualizing
in 2D the genome G along the X axis and the read Q along the Y axis, we are able
to compute  a set  of (spaced)  "anchors", which  are diagonal matching  regions
between G and Q.

In the "Window Generation" filter, we find candidate matching windows CMW, which
are   regions of G  of <window_length>  length that  contain "enough" anchors to
justify further    investigation. There  are several  window    generation modes
available (selected by -n).

In  window generation mode 2  (the  default), each CMW  must  be supported by at
least 2 anchors. In mode 1, CMW might be created even from a single anchor. Mode
1 results  in higher sensitivity, but lower  speed. In our experience, requiring
more than 2 anchors for each CMW no longer provides justifiable speed benefits.

In paired mapping mode, filter 1  is run independently for each  read in a pair.
In this case,   in window generation  mode  "-n 4" (the  default),  both runs of
filter 1 occur  in (unpaired) mode  2. In window  generation modes "-n 2" or "-n
3", both runs of filter 1 occur in (unpaired) mode 1.

The <window_generation_threshold> parameter specifies the minimum score that has
to be achieved by  an optimistic estimate of the  real match score in order  for
the CMW to be generated. The "optimistic" estimate consists of filling the holes
in any anchors, and constructing a match from  at most two  anchors and a single
gap. (Note,  we say "optimistic"  because the anchors  come from matching spaced
kmers, and G and Q might in fact differ at locations that are in those holes.)


Filter 2: Vector SW/Ungapped Alignment
--------------------------------------

Next,  all  CMWs  for   a given read   are  investigated  by a  fast  vectorized
implementation   of  the  Smith-Waterman  dynamic  programming  string  matching
algorithm. When running  in colour space  mode, the  matching in  this filter is
applied to the  colour space representations of  the  read and  the CMW, and  no
crossovers (colour space sequencing errors) are called. When running in ungapped
mapping mode,   this filter  is  replaced  by a   linear time  optimal alignment
algorithm.

In order  for a CMW to  pass this filter,  it must achieve a  score of  at least
<vector_sw_threshold>. In colour space, this is set by  -v. In letter space, the
vector SW and  full SW should obtain the  same score, so <full_sw_threshold>  is
used.

Since filter  1 can potentially  create many overlapping  CMWs that are off by a
few bp, filter 2  offers the option to  discard any CMWs  which overlap a "good"
CMW (which already passed the filter) by more than <window_overlap>. This is set
by -l.

To speed up the mapping of a single read against repetitive regions of a genome,
every time the SW algorithm  is run, the  resulting score is  saved in a  cache,
indexed by the hash value of the  genomic window. Subsequently, prior to running
the SW algorithm on another genomic window (and the same  read!), its hash value
is   computed and the cache   is looked up  for  a previous score. Conflicts are
possible, resulting in wrong SW scores, but they are extremely rare. The -Z flag
disables this mechanism.  This is disabled anyhow in  ungapped mode, as in  that
case, computing the hash value of a genomic  window is not  much faster than the
linear time alignment algorithm.


Filter 3: Scalar (Full) SW Alignment
------------------------------------

Finally,   the  remaining <num_outputs>+10 top  CMWs   for  each read are passed
through a full SW alignment  algorithm. While the  vector  SW applies to  colour
space, the  final full alignment   is always done   in letter space. Since  each
letter depends on the previous letter and colour, any  error on the colour space
read will affect all following letters when converting to letter space. For this
reason, we perform the alignment of all four  possible letter space translations
of   the   read    and permit    jumping between  matrices,    at  the   cost of
<crossover_score>.  A     CMW passes this   filter if   its score is    at least
<full_sw_threshold>.  This is set by   -h. At  the  end,  the top  <num_outputs>
remaining CMWs are output.

The full SW alignment is the costliest of the filters applied, with running time
proportional to the product  of the lengths  of the read  and  the CMW.  A small
optimization is achieved by restricting the dynamic programming to run through a
series of  "necks" or diagonal  tubes  centered along the  original anchors that
triggered the creation of the CMW in filter 1. By  default, gmapper uses anchors
with total extra width 8. In ungapped mode, the extra width is set 0. To disable
this restriction and allow  the SW algorithm to run  through the entire  matrix,
use "-a -1".


Paired Mapping Mode
-------------------

In paired mapping mode, an additional conceptual filter is added between filters
1 and  2. This filter discards  any CMWs for one read  in a pair  that cannot be
"paired up" with  a CMW for the  other read. Here, a  CMW for read Q1 is "paired
up"  to a CMW  for read Q2 if  the size of the  insert between them falls in the
allowed interval.

The final mapping score for a  pair equals the sum  of the mapping scores of the
individual reads. (We do not currently score  insert sizes.) Following filter 3,
the top <num_outputs> remaining paired hits for each pair are output.


Mapping Quality Values
----------------------

As of v2.2.0, SHRiMP computes by default quality values for mappings, as well as
for individual base calls. In letter space, base  quality values are the same as
the input base quality values. In colour space,  base quality values are derived
by running  a  forward-backward  algorithm in order    to compute the  posterior
probability of every base call.

Mapping quality values are  based  on  the  probability  a  certain  mapping  is
correct.  For unpaired reads,  the probability a certain mapping is  correct  is
derived as the posterior of that mapping divided by the sum of the posteriors of
all other mappings of that read. For reference, see:  Li H., Ruan J., Durbin R.,
"Mapping short DNA sequencing reads and calling variants using  mapping  quality
scores" doi: 10.1101/gr.078212.108.

For paired reads, a more elaborate computation is performed. Mappings are placed
into three classes:  paired, unpaired for the first  read, and  unpaired for the
second read.  Various weights are used  to carefully select between the classes,
taking into account the posteriors of the mappings, the read lengths, as well as
prior  assumptions on the sensitivity  of SHRiMP. In the  end, of all classes of
mappings  for a  certain pair  of  reads, at  most one  will have quality  value
greater than or equal  to 4.  When --single-best-mapping  is used, only one best
mapping is output for each pair.  (Note, the  quality of this best mapping might
still be  less  than  4. If  those  are  undesirable,  they should  be  filtered
downstream from the SAM output file.)

When using --single-best-mapping, if SHRiMP decides that the  best mapping for a
pair of  reads is a  singleton  mapping (only one  of  the reads being  mapped),
SHRiMP will also look at  the best singleton mapping of  the other read. If that
one is by itself strong  enough, SHRiMP will  forcefully pair up those mappings,
and mark them as an "improper" mapping. (Specifically, bit 0x2 in the FLAG field
of the SAM output will be 0). Such mappings do not respect any assumptions about
the insert size distribution, and they can even be on different chromosomes.  To
disable this behaviour, see --no-improper-mappings.


--------------------------------------------------------------------------------
7. Comparison to SHRiMP1
--------------------------------------------------------------------------------

7.1 Algorithms
--------------

The major design   difference between SHRiMP1  and  SHRiMP2 is  that the  former
indexes reads and runs  through the genome, while the  latter indexes the genome
and runs through the reads. Because of this, to make  SHRiMP2 run a machine with
limited  amounts    of RAM,  one   needs  to   split   the  genome into  several
pieces. However, the new design comes with a number of positives: genome indexes
can  be reused,  massive  parallelism  can  be achieved, and  mate-pair/pair-end
information can be  effectively used to  dramatically restrict  the search space
for mapping every read.


7.2 SHRiMP Output Format
------------------------

SHRiMP2 provides SAM output natively. For backward compatibility and for some of
the  tools included,  it still  provides the  original SHRiMP output,  described
below.

Lines corresponding   to individual hits  with  tab-delimited fields. Such lines
always begin  with a '>' character in  the first position.  All utilities ignore
any lines that do not begin with '>', such as alignments, comments, etc.

Here's an  example ('\' indicates  continuation of the  same logical line on the
next line of this README file and does not appear in the actual output):

    >947_1567_1384_F3       reftig_991      +       22901   22923   3       \
    25      25      2020    18x2x3

Additionally, the beginning of  each output file begins  with a specification of
the tab-delimited fields. For example,  the following applies  to the above read
hit:

    #FORMAT: readname contigname strand contigstart contigend readstart readend\
     readlength score editstring

The #FORMAT: line allows new fields to be unambiguously added, or others removed
or reordered without requiring alteration to the parsing routines.

Descriptions of the columns are as follows:
	'readname'	Read tag name
	'contigname'	Genome (Contig/Chromosome) name
	'strand'	Genome strand ('+' or '-')
	'contigstart'	Start of alignment in genome (beginning with 1, not 0).
	'contigend'	End of alignment in genome (inclusive).
	'readstart'	Start of alignment in read (beginning with 1, not 0).
	'readend'	End of alignment in read (inclusive).
	'readlength'	Length of the read in bases/colours.
	'score'		Alignment score
	'editstring'	Edit string: read to reference summary; see below.

Additionally, probcalc output adds the following columns:
	'pchance'	Probability that a read will align with a genome with 
			as good a score or better by chance.
	'pgenome'	Probability that a hit was generated via common 
			evolutionary events characteristic of the genome.
	'normodds'	Normalized pgenome/pchance.

The  'editstring' column contains  a  short  description  of the  alignment with
reference   to  the  database  sequence.   This  allows at-a-glance analysis  of
alignments, including identification of miscalls, SNPs, indels, etc.

The edit  string consists  of numbers, characters   and the following additional
symbols: '-', '(' and ')'. It is constructed as follows:
    <number> = size of a matching substring
    <letter> = mismatch, value is the tag letter
    (<letters>) = gap in the reference, value shows the letters in the tag
    - = one-base gap in the tag (i.e. insertion in the reference)
    x = crossover (inserted between the appropriate two bases)

For example:
    A perfect match for 25-bp tags is: "25"
    A SNP at the 16th base of the tag is: "15A9"
    A four-base insertion in the reference: "3(TGCT)20"
    A four-base deletion in the reference: "5----20"
    Two sequencing errors: "4x15x6"	(i.e. 25 matches with 2 crossovers)


7.3 SHRiMP1 Tools
-----------------

Some of the tools included with the original SHRiMP1 package are still available
(probcalc,  prettyprint,  shrimp_var),  but  these tools do  not   work with SAM
output.

For  a description of these  tools, we  include  the README file  for the latest
SHRiMP1 version, in README.SHRiMP132.


--------------------------------------------------------------------------------
8. Contact
--------------------------------------------------------------------------------

The program website is 
    	http://compbio.cs.toronto.edu/shrimp

The authors of this software may be contacted at the following e-mail address:
	shrimp at cs dot toronto dot edu


--------------------------------------------------------------------------------
9. Acknowledgements
--------------------------------------------------------------------------------

Development  was performed at the University  of Toronto's Computational Biology
lab in collaboration with the Stanford University Sidow Lab.

The   development of this   distribution was made possible  in  part by National
Engineering and Research Council of Canada Undergraduate Student Research Awards
(NSERC USRAs).

We would like to thank Dr. Lucian Ilie of the  University of Western Ontario for
providing  us  with sets  of spaced  seeds  especially designed  to achieve good
sensitivity.

We would  like to thank  Dr. Alessandro  Guffanti (Bioinformatics, Genomnia srl,
Milan,  Italy),   Christine Voellenkle and Jeroen  van   Rooij ( Policlinico San
Donato, Milan, Italy) for their  feedback on using  SHRiMP for mapping micro RNA
data, including the set of spaced seeds mentioned in section 3.5, which are used
by default in miRNA mode.